1.2 KiB
ADR 0029 — NVML truth and multi-resident placement
Status: accepted
Context
The M5–M9 scheduler safely serialized one resident deployment, but that prevented proven small capabilities from coexisting. Naively subtracting both NVML-used and managed residency would double count memory; trusting absent telemetry as zero would overcommit the GPU.
Decision
NVML used is physical truth. Managed residency is attribution within that observation, and only unmaterialized lease bytes are additionally reserved. Unknown/stale telemetry fails closed. A single versioned planning function drives dry-run and actual admission under a database accelerator lock. The runtime worker may retain multiple exact deployment identities. Eviction is limited to managed idle residency and recovery adopts only an exact immutable identity report.
Consequences
GPU Node can serve multiple capabilities when current observations and measured envelopes prove fit, without weakening the 1 GiB minimum reserve. External workloads remain visible but unmanaged. Conservative rejections may reduce utilization during telemetry uncertainty or pressure; that is an intentional safety result, not a capability-health failure.